Using AI in problems that are easily verifiable is a good way to save time without falling victim to hallucinations.
PCB review is a great application. Having the AI review your board at the same time you are can catch extra mistakes with pretty much no drawbacks.
A false positive has 0 cost (other than wasting your time a bit), and can be easily verified.
A false negative just means the AI misses a mistake. It doesn’t make your review worse.
The AI is able to go over details at a superhuman level and check things that a human would get bored by.
Another great application is BOM generation. You can easily check the part descriptions in JLC against your Value fields.
By the same principle, some problems are ill suited. Footprint generation is very difficult to verify.
You need to measure pretty much every single geometry, which is pretty much the same amount of work as doing it manually.
And I’m still on the fence about AI-generated software testing.
It isn’t like a PCB review which just gets thrown away, the tests stick around in your codebase,
and mistakes can lead to false positives that waste a bunch of time because it isn’t obvious whether the test or the code is following the true spec.